ExoMind

ExoMind: Democratizing Scientific Intelligence via Extended-Mind-Inspired Agentic System

Surpassing Frontier Proprietary Models in Scientific Reasoning and Research with Less Data, Small Model, and Low-cost Training

ExoMind Team Shanghai Artificial Intelligence Laboratory

ExoMind is the first extended-mind-inspired agentic system, integrating systematic data engineering, a scientific interaction framework, and a systematic training strategy to advance scientific reasoning and research.

Technical ReportPDF ↗ GitHubRepository ↗ Hugging FaceModel ↗ ModelScopeModel ↗

Frontier-level scientific intelligence at 35B scale

ExoMind consistently improves over the base model on all eight datasets and surpasses leading open and closed-source models on most of them, achieving frontier-level scientific intelligence at 35B scale with roughly 277× fewer parameters than GPT-5.5.

+14.5

Average benchmark score vs. GPT-5.5 (xhigh)

~277×

Model-size efficiency vs. GPT-5.5 (xhigh)

~10K×

GPU-count efficiency vs. GPT-5.5-scale training

~100K×

Training-data efficiency vs. GPT-5.5-scale training

Scientific intelligence evaluation comparing ExoMind with frontier and open models

Intelligence beyond a single LLM

Scientific intelligence demands fine-grained, long-horizon reasoning that is difficult to compress into a single model. Inspired by the Extended Mind Thesis, ExoMind organizes intelligence as an agentic system of an LLM, typed interaction objects, and autonomous interaction processes.

A broader paradigm for scaling intelligence

Deep specialization

Shift specialization from repeatedly modifying model parameters toward designing interaction objects that can be tailored to the operations of a task or domain.

New scaling dimensions

Expand capability through richer interaction objects and deeper autonomous interaction, beyond model size, data volume, and training time.

Stronger representational capacity

Organize the model, external objects, and interaction processes as one system for diverse, fine-grained, and long-horizon tasks.

Building ExoMind

ExoMind is built through three connected systems rather than a single training recipe: data engineering determines what is worth learning, interaction design determines how scientific work is performed, and progressive training aligns the model with that process.

ExoMind starts with a large and diverse scientific problem-answer pool, uses strong models to predict each problem's difficulty and potential tool benefit before trajectory generation, and progressively filters and routes problems into pure-reasoning or interaction-reasoning paths. This yields a compact set of high-quality data with stable learning signals.

Scientific and general intelligence evaluation

Evaluate ExoMind along two complementary dimensions: scientific intelligence against frontier open and proprietary models across eight benchmarks, and general intelligence against the 35B base model across six benchmarks.

Benchmark explorer

What the results show

Scientific research

ExoMind reaches 70.0 on FrontierScience-Research and 84.0 on CMT-Benchmark, substantially exceeding the strongest external results, while remaining competitive on the highly challenging HLE and CritPt benchmarks.

Scientific reasoning

It leads the compared models on AMO-Bench, IMO-AnswerBench, HiPhO, and FrontierScience-Olympiad with 78.0, 92.8, 49.7, and 89.0.

General intelligence

ExoMind improves over the 35B base model on all six general benchmarks, including MMLU-Pro, GPQA-Diamond, MedMCQA, LiveCodeBench, xbench, and τ2-Bench, where the score rises from 32.5 to 74.8.

Evaluation scope

The evaluation covers frontier proprietary and open models, eight scientific reasoning and research benchmarks, and six complementary general intelligence benchmarks.

Incompressible knowledge probe evaluation

IKP evaluation shows that ExoMind substantially improves the IKP accuracy of the base 35B model and reaches a level close to models with approximately one trillion parameters, suggesting a new dimension for scaling scientific intelligence through interaction.

IKP evaluation showing ExoMind at 35B approaching the capability range of much larger models

Emergent scientific interaction patterns

More available interaction objects do not automatically produce stronger performance. The key is whether the model can organize finite atomic capabilities into an observation-guided process: identifying its current uncertainty, selecting an appropriate interaction object, absorbing the returned observation, and updating subsequent reasoning. Through hybrid progressive CoI training, ExoMind composes these objects into reusable verification-seeking, evidence-seeking, and hybrid closed-loop patterns.

FrontierScience-Olympiad #04

Resolve a sign ambiguity through executable verification

1.0 / 1.0
Research question

Derive the electric field outside a grounded conducting cylinder in a uniform transverse field and verify that the result satisfies the physical boundaries.

Uncertainty
A sign-sensitive derivative

The cylindrical-gradient factor makes the sign of the tangential field fragile.

Interaction
Code Executor

Differentiate the potential symbolically instead of trusting manual recall.

Observation
Both components are returned

The executable result exposes the exact radial and tangential signs.

Reasoning update
Verify physical limits

Confirm the surface boundary and recover the applied field at infinity.

Trajectory evidence
Reasoning state

Isolate the fragile step

The model starts from a grounded-cylinder potential that satisfies the boundary and far-field conditions, then identifies the tangential derivative as the step most likely to contain a sign error. 

Phi = -E0*(r - a**2/r)*cos(theta)
Etheta = -(1/r) * dPhi/dtheta
Action Code Executor

Differentiate in cylindrical coordinates

Phi = -E0*(r - a**2/r)*sp.cos(theta)
Er = -sp.diff(Phi, r)
Etheta = -(1/r)*sp.diff(Phi, theta)
sp.simplify(Er), sp.simplify(Etheta)
Observation Code Executor

Executable feedback fixes the sign

Er = E0*(a**2 + r**2)*cos(theta)/r**2
Etheta = E0*(a**2 - r**2)*sin(theta)/r**2
State update

Absorb the result into the derivation

Rewriting the returned tangential component confirms the negative sign. ExoMind then checks that the tangential field vanishes at r = a and that the field approaches E0 x-hat as r tends to infinity.

Related scientific intelligence projects

Scientific reasoning models, benchmarks, and multi-agent systems from the broader open ecosystem.

P1

Physics olympiad reasoning models trained with reinforcement learning.

HiPhO

High-level physics olympiad benchmark for multimodal scientific reasoning.

JiSi

Routing and aggregation system for scalable model selection and evidence use.

Citation

Please cite the ExoMind technical report using the following entry.

BibTeX 
@misc{exomind2026,
  title  = {ExoMind: Democratizing Scientific Intelligence via Extended-Mind-Inspired Agentic System},
  author = {Peng Ye and Zhuo Liu and Jingqi Ye and Fangchen Yu and Shengji Tang and Yichen Jiang and Haonan He and Zongsheng Cao and Tao Chen and Bo Zhang and Wanli Ouyang and Bowen Zhou and Lei Bai},
  year   = {2026},
  note   = {Technical report},
  url    = {https://github.com/AI4SGI/ExoMind/blob/main/Paper.pdf}
}
Expanded research figure